Guide

How to spot agentic use cases in your company

Not every process needs an AI agent. This guide gives you a method to decide, process by process, whether an agent fits, whether an AI workflow or a rule is enough, or whether it is better not to automate yet.

In short

Place each process on the lowest rung of the ladder that solves the problem. An agent fits when the steps cannot be written in advance, the tools and data are accessible, and someone can verify the result. If any of those is missing, the answer is usually a rule, an AI workflow, or waiting.

The automation ladder

Every process lands on one of five levels. The default recommendation is the lowest level that solves the problem, and you only move up when you need to.

Level 0 Do not automate yet
What it is
Redesign, measure, or leave the process as it is.
When it fits
Low volume, a broken process, no owner, or no baseline.
Warning sign
Automating a broken process only speeds up the error.
Level 1 Rules and traditional software
What it is
Deterministic logic, forms, integrations, RPA.
When it fits
Fixed steps, written rules and structured inputs.
Warning sign
If a person has to interpret free text, rules are not enough.
Level 2 AI workflow
What it is
A language model inside predefined steps: extract, classify, summarize, draft.
When it fits
Predictable steps with unstructured inputs, such as emails, PDFs or open-ended forms.
Warning sign
If the next step depends on what is found, a workflow falls short.
Level 3 Agent with human approval
What it is
The model decides its own steps and uses tools, and a person approves before anything takes effect.
When it fits
Steps that cannot be foreseen, tools available and costly errors.
Warning sign
Without tools or accessible data, the agent has nothing to work with.
Level 4 Autonomous agent with controls
What it is
The agent acts on its own within limits, with logging and automatic verification.
When it fits
Open-ended steps, high volume and results that can be verified automatically, or cheap errors.
Warning sign
If nobody can verify the result, it must not be autonomous.

The questions that identify an agentic case

Ask these questions about each process. Group A measures value, group B checks the minimum conditions, and group C tells a workflow from an agent. “I don't know” is a valid answer: write it down and keep going.

A. Value

Is this process worth touching?

How often does it happen, and how many people do it?: What it decides Whether it is worth automating.
What does an error cost, and who catches it?: What it decides Whether human approval is needed.

B. Conditions

Two of these questions are vetoes. If the answer to either one is no, the process stays at level 0 until that is fixed.

Is there a named owner who answers for the result?: What it decides Veto. Without a process owner, it is not automated yet.
Are time, volume or errors measured today?: What it decides Veto. Without a baseline there is nothing to compare the result against, and no way to pay by outcome.
Is the input data digital and reachable by export or API?: What it decides Whether it is feasible.

C. Workflow or agent

With value and conditions clear, these questions pick the level.

Can the steps for the typical case be written down in advance?: What it decides If yes, level 1 or 2.
Does what to do next depend on what is found in each case?: What it decides If yes, level 3 or 4.
Which systems or tools would whoever does the work have to use?: What it decides Whether an agent is viable.
How do you know the result is correct: by a rule, by a person, or there is no way to know?: What it decides Whether it can be autonomous (level 4) or needs approval (level 3).
How fast must it respond, and how much can each case cost?: What it decides Whether the cost and latency of an agent are justified.

Examples by area

Inside the same area, each task lands on a different level. These examples are generic and illustrative: the right level for your process depends on your volume, your data and your controls.

Finance

Level 1: Rules and traditional software Matching bank movements against the ledger when the matching rules are already written down.
Level 2: AI workflow Reading supplier invoices that arrive as PDFs in different layouts and extracting the fields for the accounting system.
Level 3: Agent with human approval Investigating a difference that does not reconcile: the agent looks across several systems and proposes the adjustment, and a person approves it before it is posted.

Operations

Level 0: Do not automate yet A process that each shift runs in a different way and nobody measures. First agree on one way of doing it, then measure it.
Level 1: Rules and traditional software Moving an order from one system to another when the fields are structured and the steps never change.
Level 3: Agent with human approval Resolving an order exception, where the next step depends on what the agent finds in inventory, shipping and customer records, with a person approving the final action.

Customer service

Level 1: Rules and traditional software Answering an order status question with a system lookup and fixed steps.
Level 2: AI workflow Sorting incoming messages by topic and urgency and drafting a reply for a person to send.
Level 4: Autonomous agent with controls Simple, high-volume requests that an agent resolves on its own within limits, when the result is verified automatically and every action is logged.

Sales

Level 1: Rules and traditional software Assigning a new lead to a rep by territory, with a written rule.
Level 2: AI workflow Summarizing a call and filling in the CRM fields from the notes.
Level 3: Agent with human approval Researching an account before a meeting and preparing a draft proposal: the agent decides where to look, and a person reviews before anything is sent.

Compliance

Level 1: Rules and traditional software Checking that a file has every document on a fixed checklist.
Level 2: AI workflow Extracting obligations and dates from contracts or regulatory notices written in free text.
Level 3: Agent with human approval Reviewing a case against several sources, where each finding changes the next step, with a responsible person signing off. If nobody can verify the result, the agent must not be autonomous.

Human resources

Level 0: Do not automate yet An onboarding process with no named owner. First assign the owner.
Level 1: Rules and traditional software Sending the same onboarding forms and reminders on fixed dates.
Level 2: AI workflow Sorting employee requests that arrive as free text and drafting an answer from the written policy for a person to review.

Where not to use an agent

A good map also says no. An agent is the wrong choice when:

The steps can be written in advance. A rule or an AI workflow solves the case, and an agent only adds cost and latency.
The process is broken, has no owner, or has no baseline. Automating it only speeds up the error.
Nobody can verify the result. Without verification there should be no autonomy.
There are no tools or accessible data. The agent has nothing to work with.
The volume is low. The effort of automating is not justified yet.

Sometimes the answer is a rule. A map that says “a rule is enough here” is more credible than one that sees agents everywhere.

Melvin

An AI agent by HechoX

Map your company with Melvin

This guide teaches the method. Melvin, the AI agent by HechoX, applies it to your company: it walks through these questions with your processes and gives you a Map with the recommended level for each one, the three best candidates, and where not to use an agent.

Do my Map with Melvin

Free with your work email.